[nfsv4] Question about NFSv4 change attribute

[Jeff Layton <[e-mail]>]

tl;dr: if the POSIX atime (and only the atime) is updated due to a read,
do we need to update the change attribute?

Longer question...

I've been doing some work around the i_version counter in the Linux
kernel, which is used to implement the change attribute for NFSv4 for
some filesystems. What I've found is that there is some inconsistency in
the behavior between different exportable filesystems. In particular
btrfs does not bump the change attribute for atime-only updates, but
ext4 and xfs currently do.

The spec basically says "the value must change whenever the file data or
metadata changes". While an atime update is technically a metadata
change, it's not a "real" change to the inode. Changing it generally
causes cache invalidations, which we don't want to incur just because
someone did a read.

I think we probably ought to avoid incrementing the change attr when
only the atime is being changed. Am I wrong? What do other server
implementors do?

It might be nice to add an addendum to the spec that spells this out
either way.

Thanks,
-- 
Jeff Layton <[e-mail]>